Piecewise-constant parametric approximations for survival learning


Jeremy C. Weiss ;
Proceedings of the 2nd Machine Learning for Healthcare Conference, PMLR 68:1-12, 2017.


Logged events occur both regularly and irregularly over time. In electronic health records, these events represent mixtures of scheduled and urgent or emergent encounters. Whereas most survival models use baseline events to estimate the rate function for an outcome, e.g., Cox processes using the proportional-hazards assumption, our framework uses logged events over time to predict survival outcomes with piecewise approximations of arbitrary hazard functions. We develop a procedure to learn forests as combinations of piecewise-constant and parameterized distributions to compactly model survival distributions from data. Under this construction, the model provides a “now-time” risk that incorporates irregularly-repeated data and for health outcomes serves as a surrogate for patient disposition. We illustrate the advantages of our method in simulations and in longitudinal, intensive care unit data of individuals with diabetes admitted for ketoacidosis.

Related Material